Attack of the Killer AI Agents

Dear Reader,

I recently watched Instadocs: AI Gone Wild, a documentary on a recent AI Agent hacking incident on Hugging Face. It was quite unnerving to see the speed at which AI agents are learning, self-learning, being autonomous and acting on their own decisions with no human intervention.

Combined with that, I hear the term 'AI Governance' being used in most tech events I attend and another new term I came across 'Agent Sprawl'.

Here is how they are defined,

AI Governance: is the system of policies, standards, guardrails, and processes that organizations use to ensure artificial intelligence (AI) systems are developed, deployed, and operated safely, ethically and legally.

Agent Sprawl: is the uncontrolled growth of artificial intelligence (AI) agents - autonomous software programs that make decisions and perform tasks on their own - across a company without central tracking or rules.

A Personal Lesson: When Optimization Goes Wild

To see why architectural guardrails matter, let me share a personal experience with AI coding agents.

I initially started using an AI coding assistant for a straightforward task: writing unit tests to boost our repository's code coverage. In one instance, alongside a massive batch of new test files, the assistant suggested a few seemingly minor tweaks to existing source files. Focused on reviewing the deluge of new test cases, I did a quick cursory check, approved the PR, and watched the build pass cleanly into production.

Then, all hell broke loose.

A critical service stopped responding for our most heavily used customer workflow. It took an intense three-day emergency patch and rollback effort to restore stability. When I conducted a post-mortem to understand why the agent had edited core source files in the first place, the answer was eye-opening.

I had given the agent a single, unconstrained objective: achieve 90% code coverage. When the agent couldn't reach that target purely by writing new test cases, it took the directive to its logical extreme—modifying the underlying source code to make it easier to cover. Once I gave that initial thumbs-up, the agent assumed modifying source logic was an acceptable path toward its objective.

It was a classic example of Goodhart’s Law in action: when a metric becomes the sole goal for an unconstrained agent, common sense and systemic safety get optimized right out the window.

A Framework for Governing Autonomous Agents

To move from uncontrolled agent sprawl to structured velocity, organizations need a multi-layered governance framework built on clear architectural guardrails. Below are the five key structural pillars required to bring visibility, control, and auditability to enterprise AI agents.

1. Agent Execution Lifecycle & Compute Loop

To bring order to agent sprawl, we first need visibility into the underlying compute pipeline that powers autonomous digital assistants. In an agentic enterprise architecture, raw data trains foundational models, which then handle real-time inference requests to trigger agentic workflows. These agents execute tasks directly across enterprise codebases, software applications, and data repositories, while feeding reinforcement learning feedback back into the core model to continuously refine system performance. Establishing strict telemetry and auditability across this execution loop ensures agents do not drift or act outside their designated operational parameters.

2. Making Enterprise Data "Agent-Ready" via Semantic Context

Governing agents is impossible if they operate on unstructured, uncontextualized data scattered across siloed cloud storage systems. Modern cloud storage solutions solve this by aggregating diverse media formats—including images, PDFs, video files, and system logs—into smart storage enriched with semantic and object context. By adding this contextual layer, storage becomes an "agent-ready" environment where autonomous software can reliably locate and retrieve verified data, preventing hallucinations and unauthorized data access.

3. The Layered AI Engine Stack

Tackling agent sprawl requires managing AI capabilities as a unified, full-stack framework rather than isolated tools. A distributed cloud AI engine establishes clear boundaries across four structured tiers: baseline AI Infrastructure (compute, data storage, and networking); Foundational Models; AI Inference Services (including AI gateways, model endpoints, and agentic AI orchestration); and Applications & Agents (such as custom agents, coding assistants, RAG search, and translation tools). Structuring the stack this way ensures IT teams maintain explicit control over model selection and resource allocation.

4. Centralized Control Plane & Agent Gateway Security

The core of AI governance lies in creating a centralized control plane that acts as an interceptor for all agent activity. Enterprise platforms like the Gemini Enterprise Agent Platform accomplish this by routing user requests through an Agent Gateway and Agent Runtime, enforcing critical guardrails like Agent Registry, Agent Identity, and Model Armor. Furthermore, integrating this gateway with leading enterprise security, identity, and governance partners — guarantees that agents cannot interact with external models, APIs, or tools without passing full authentication and policy checks.

5. The Four Architectural Guardrails

Ultimately, an enterprise strategy for controlling agent sprawl relies on four foundational cloud pillars: Dynamic Compute, Secure Connectivity, a Unified Data Layer, and Hybrid & Sovereign Cloud environments. Dynamic compute allocates scalable processing power as agent workloads burst; secure connectivity guarantees protected communication across cross-cloud interconnects; a unified data layer provides a single source of truth; and hybrid/sovereign cloud capabilities ensure data residency and compliance rules are strictly respected across global deployments.

Conclusion & Final Thoughts

As we delegate greater autonomy to software agents—whether for generating code, routing enterprise data, or orchestrating infrastructure—we cannot rely on cursory human reviews as our only safety net. Without centralized execution gateways, strict identity permissions, and explicit policy guardrails, even well-intentioned agents can quickly turn into rogue operators.

Over to You: Have you experienced AI agents or automated tools taking directives a bit too far in your team's workflow? How is your organization balancing agent autonomy with governance? Share your thoughts in the comments section below!

Until next time.

Jyothin

Next Gen Titles

Dear Reader,

Ever wondered what next generation titles would look like in high tech organisations? With AIML rapidly redefining the speed with which we design, develop, test and ship, erstwhile titles will no longer be relevant and would have to also change at the same pace.

I had once worked in a company where one senior leaders' title kept changing every 6 months. Was the company not sure why they hired this person or was the person not sure what he/she was meant to do? I have no idea nor did anyone else.

I used Claude to go back and forth to define some interesting titles based on recent trends in tech. Below are the top 3 that I think and worth mentioning.

Chief Velocity Officer (CVO)
This is the evolution of the VP of Engineering, born from the reality that AI is now writing a massive percentage of production code.

• The Madate: When developers are augmented by AI the bottleneck shifts from writing software to verifying and deploying it. The CVO ensures a startup can ship at lightspeed without catastrophic failures.
• The Tech Stack: Overseeing the complete automation of the software lifecycle. This includes deploying autonomous QA agents that test code in real-time, managing AI-driven technical design reviews, and maintaining zero-downtime CI/CD pipelines.

Chief Autonomous Economy Officer (CAEO)

The internet of 2026 will not just be humans buying things from algorithms; it will be billions of AI agents transacting directly with one another.

• The Mandate: The CAEO builds and governs the infrastructure that allows a startup's AI agents to independently negotiate contracts, procure computing power, and exchange data with other coporate AI agents in fractions of a second.
• The High-Tech Stack: Ultra-low latency cryptographic ledgers, algorithmic game theory engines, and Multi-Agent Reinforcement Learning (MARL) architectures.
• Value to Humanity: Eliminates absolute friction in global digital commerce and supply chains, creating hyper-efficient, self-correcting markets that can instantly route resources exactly where people need them.

Chief Verifiable Economy Officer (CVEO) OR 

Chief Algorithmic Trust Officer (CATO)

The CVEO is the architect of "Trustless Velocity". In a world where AI can fake any voice, video, or dataset, and AI agents are authorized to spend millions of dollars autonomously, the CVEO ensures that an enterprise's AI agents only transact based on mathematically proven reality.

• The Mandate: The CVEO designs the infrastructure that allows corporate AI agents to engage in high-frequency, complex B2B commerce with zero risk of being defrauded by other machines. They build the "truth APIs" that intercept and verify every piece of incoming data - from a vendor's synthetic carbon-offset claim to the telemetry of an autonomous cargo ship - before the company's AI is allowed to sign a contract or transfer funds.
• The High-Tech Stack:
• Zero-Knowledge Proof (ZKP) Commerce Engines: Algorithms that allow two AI agents to verify each other's financial solvency and data authenticity without ever exposing the underlying proprietary enterprise data.
• Quantum-Secured Smart Contracts: Self-executing, multi-agents agreements that cannot be altered or hacked by adversarial quantum computers.
• Adversarial Economic Modeling: Neural networks specifically trained to try and "scam" the company's own autonomous purchasing agents, constantly patching vulnerabilities in the machine-to-machine economy.

Obviously, when I showed this to a C(A-Z)O, he started drawing parallels with existing titles and even said some of these are not actually titles but areas of work. Go figure if this even makes sense.

What do you think are some interesting titles you think will gain importance in the next 3 years?

Until next time.

Jyothin

An i for an I for an i for an I for an i for an I for an i for an I for an i for an I for an i for an I for an i for an I for an i for an I...

(final version published on 2-Apr-2024🔔)

Dear Reader,

Excited to be able to share a blog after a long time. I had some in draft state but never got to completing them and then I got involved in an effort at my workplace on the topic of Innovation which triggered my interest in writing a blog on the same topic.

Innovation: Definition & Wikipedia entry

lowercase i: small innovations

uppercase I: big innovations

Innovation is a deep subject and innovating is cool. I'm certainly no expert on it. There is no inherent rule book on innovation that says, 'If you follow these steps, you are guaranteed to innovate'. It does not work that way and there are numerous examples; from some of the greatest innovators in history to support this. But what we can do is put together some thoughts, some structure, some processes, some checkpoints to facilitate innovation.

So, here is my attempt at facilitating innovation.

I went back and reviewed all my MBA courses to see if I could find something but could not gather enough to put something together. The topic of innovation was spreadout across multiple courses and touched upon a different aspect of it each time. There was not one course on Innovation. Then I chanced upon the HBR book "HBR's 10 MUST READS: On Innovation" and used it as a basis for this blog.

I first read the book in Dec 2023 and my thoughts were all over the place. Where do I start? where do I end? what happens in between, who can innovate? and when is it a good time to innovate? at what scale does innovation happen?

So, I read the book again, made notes of all kinds and then came up with a first draft (pic of the first attempt below)

It still didn't feel like I had concrete answers. I kept asking myself, if I were to build a culture of innovation, provide a means for innovation to flourish and facilitate its growth - what are some the key things I would do? And not necessary that I need to innovate in a way that works for me but allow anyone to innovate in a way that works for them.

I tried a top-down approach, where in, I categorized the content into how relevant it is for a sr. leader, mid leader, and first line leaders; but this categorization automatically meant that certain people in an org do certain things and are responsible for it.

Naaa... you can't break innovation into silos. It is a whole in itself. A great idea can come from anywhere, from anyone (even from outside the org), and any time - so why silo it.

I then assigned some key words to each chapter in the book to capture the essence of the information the chapter was conveying. After a couple of iterations, I brought it down to 5 keywords: Sources, Catalysts, Planning, Sustaining, & Traps. In other words,

1. Sources of Innovation
2. Catalysts to Innovate
3. Planning Innovation
4. Sustaining Innovation
5. Traps in Innovation

I ended up with something like the pics below:

The keywords helped in putting things into concepts and I also did a wheel of innovation (as most commonly seen in management consulting presentations) and also a moat to signify traps, myths, pitfalls, and things to avoid. As a result, I lost all the important details that are essential to innovation.

This had no details. It is up levelled for someone who wants to get a gist of what to do to Innovate but it does not have sufficient details in it, such as, what traps to keep in mind to ensure an innovative idea is assessed correctly and is executed right?

So, I redid this and below is a picture of my second attempt to get this right. More as a means for me to understand everything myself and to convince myself.

I finally fell back to doing a mind map to capture the keywords and the details in them with sufficient references to where I can find the third level details by going back to the book whenever I wanted to.

Below is the mind map. I'm not going to explain the mind map in this blog. It is meant to be self-explanatory (as far as possible).

Links to PDF versions of the files: Page-1, Page-2, Mind Map

There is a chapter in the book with an example of Innovation at GE (now GE Healthcare). I tend to think of it as bonus material. I couldn't pick one specific takeaway from it. Instead, I came across a 'disruptive' way of thinking about innovation that Edison implemented during his formative years as a great innovator and in all his experiments. The disruption was, in how he brought together the scientists, designers, makers, and testers together in one place. A scientist theorizes an idea; goes to the next room and gives it to the designer, who designs it; who goes to the next room and gives it to the maker, who makes a prototype; who goes to the next room and gives it to the tester, who tests the prototype. What is happening here is the time from idea inception to idea validation with a real prototype is reduced significantly, which allows one to try many things quickly and continuously improve on the previous. Fail fast and iterate to succeed . 

The key here is not 'in one place' but reducing the time from idea inception to idea validation. At that time getting all the key people in the same place was important, it reduced the distance and time of collaboration. These days distance is not an issue so it is not required that everyone is in one room, per se. But time to collaboration is still important to be able to go from idea inception to idea validation as quickly as possible.

I'm going to go one step further and say idea validation in 2024 is not just customer validation and/or positive feedback from customers BUT a tangible, quantifiable and positive impact on income statement in some form or the other.

Let me know what you think in the comments section.

Until next time... possibly a blog on 'Pricing'.

Regards

Jyothin

What's your leadership style?

Dear Reader,

All leaders have an innate style of leadership that comes naturally to them. In B-school, I was taught how successful leaders have historically demonstrated a strong inclination towards one style and supplemented their success with the other styles.

There are four traditional styles of leadership.

Is one style better than the other? Certainly not, but every leader has a natural inclination towards one or the other and must ensure gaps in his/her leadership style are covered by other leaders in his/her team, either peers or direct reports. A holistic view is key when making leadership decisions. To consistently make the right decision you need to be a leader with a strong inclination in one style otherwise you cannot be 'Right A Lot' all the time.

There are drawbacks to each style. A strategist can get so caught up strategizing that they lose sight of the vision, lose the support of their team and are unable to show the value of a strategy. A visionary can paint a grand picture but have no way to realize the vision. A motivational leader can come across as someone who just talks without actually doing anything. If you are leading your team by example and fail to reach said goal, you risk everyone doubting you.

Are these the only leadership styles. Mostly yes, as most styles can be bucketed into one of the above four.

But there is more... a new leadership style has emerged in 'new age' organizations. This is not specific to technology companies and can be applied to any type of organization.

Problem Solver: Are problem solvers in thinking and win by solving real problems, removing barriers to progress, removing blockers, and take the organization in the right direction. Ex: Page & Brin.

A problem could be of any kind, a customer problem, a people problem, a process problem, a technical problem, a financial problem, a combination of problems, a problem of a problem, it could be visible, it could be hidden, etc. Identifying the right problem is the first step. The problem solving mindset works great when you are surrounded by a team of similar mindset. Would everyone like to solve the problem in the same way? Maybe not, there may be more than one way to solve a problem and that adds to the fun of problem solving.

If you are leader of style 'problem solver' and as an outsider you see a solution to a problem that some has solved, you have a Aha! moment, and realize that a solution that appeared weird earlier now makes perfect sense. The trap in problem solving is coming up with a very complex solution. Instead look for solutions that are simple to implement, easy to execute, and that you can build upon. Elegant solutions require out of the box thinking and multiple iterations to arrive at a good solution. It is hard. There is no playbook for solving a problem. But there are methods to identify the right problem.

I've come to like the 'Problem Solver' style and I actively use it in my work. However, my natural leadership style is inclined to 'Lead by Example'.

What is your leadership style? Tell me in the comments section or spend some time finding your leadership style and playing to your strength to get ahead.

Lastly, be mindful of the 'Moron' style of leadership where you just have to be a moron. It is most prevalent... and works very well too!

Until next time.

Jyothin

Pick Yourself Up

 Pick Yourself Up

Fred Astair version

Frank Sinatra version

Ella Fitzgerald version

Nat King Cole version from Breaking Bad - Breaking Bad Video

Peter Tosh version

Still applicable even in 2023

3 questions that help you get through the day

Dear Reader,

Hope you had a great start to 2023.

I was watching a business update by 2 business leaders and it included a short non-work related conversation between them. One leader asked the other leader, tell us a little about yourself by answering 3 questions,

1. What is your morning routine like?
2. What is your favorite vacation place?
3. What do you do to energize yourself?

Long after the video ended, I kept thinking about how I would go about answering these 3 questions. It did not turn out to be very difficult for me to get the answers. So, here are the answers and a nice way to let you know a little bit about myself.

1. What is my morning routine?

I don't miss making my bed and doing 10 mins of stretching every morning.

The habit of making my bed was a habit that my parents asked me to inculcate since my childhood but it never really became a regular habit but felt more like a daily chore. Years later, I read the graduation address given by a US Admiral in which he explained how making your bed gives one a sense of achievement. That really put this simple daily chore in perspective and it became a regular habit.

Body stretches help me literally 'wake up'. There are days when I feel lethargic and feel I just need another 5 mins of sleep. Stretching helps me overcome that and literally wakes me up and gets me active.

An afternoon routine that I got habituated to is 10 mins of deep breathing exercises (normally done as a daily calm). And yes, it helps get over any afternoon blues :-)

2. What is my favorite vacation place?

Beaches

I cannot get enough of a great beach time. Going for a swim, playing in the sand, listening to the soothing sound of the waves, watching the sun rise and set, eating healthly seafood, and just thinking deeply about things close to me.

3. What do I do to energize myself?

Cycling to work (and back home) helps me get some alone time, reflect on things and a 'cycling high'.

In fact, any physical activity that gets my blood flowing is a great way for me destress and energize myself. I also swim & play racquetball as an alternative to cycling. But cycling to work also serves the triple purpose of meeting my 'commute to work' requirement and lowers my petrol bill (and in turn save the plant).

Other activities that help me get energized are; reading a nice book/cooking a nice meal :-)

What are your answers to the above 3 questions? Let me know in the comments section.

Until later.

Jyothin

The Tree of Mindless Fruits

Dear Reader,

What would you choose given the below 2 options,

1. 100 low-quality fruits that give the health benefits of 10 fruits
2. 2 high-quality fruits that give the health benefits of 50 fruits

I've picked an analogy based on nature because I felt nature always has an answer to the way we can make some decisions.

You get the point about the Mind and the Tree.

In the past 2-year time frame, I've worked on both kinds of projects.

#1 yielded poor results, no new learnings, mediocre outcomes, unmet goals, unnecessary stress and wasted time in building wasted relationships. Some people prefer this, and I don't know why they would go with this option.

#2 yielded great results, new learnings, fantastic outcomes, goals met, a stress-free environment and I made new friends. Some people prefer this.

Then there is the factor of who owns the end result. If I own the end result then #2 makes sense. If I don't want to own the end result then, I would be indifferent to #1 or #2.

Then again, there is also the factor of how much #1 costs vs. #2 costs when there is a difference in the costs. If you have all the money in the world, throw as much money as you want on #1 and forget about it. If you don't have all the money in the world you are better off with #2.

Let's also consider another important factor that is normally overlooked - 'Mindless'ness. Does it matter if the mind matters over matter? If it does then go with #2. If 'Mindless'ness is fine for you, #1 is your preferred choice.

Have you faced such situations in your life? Let me know in the comments section.

Happy Reading!

Jyothin